Source code for scitex_agent_container.config._provider_types

"""ProviderSpec dataclass for ``spec.claude.provider`` (backend override).

Lives in its own module to keep ``_types.py`` under the project's
512-line cap (mirrors ``_proxy_types.py``). Re-exported from
:mod:`scitex_agent_container.config` alongside the rest of the spec
dataclasses.

Also home to :data:`AgentProvider` (``spec.provider``, TOP-LEVEL) — see
the naming-collision note below before touching either symbol.
"""

from __future__ import annotations

from dataclasses import dataclass, field
from typing import Literal

# ---------------------------------------------------------------------------
# Agent SDK family selector — spec.provider (TOP-LEVEL; openai-compat-1
# foundation, scitex-todo card)
# ---------------------------------------------------------------------------
#
# NAMING COLLISION WARNING: this module ALSO defines :class:`ProviderSpec`
# below, for the pre-existing ``spec.claude.provider`` (nested) — a vendor-
# agnostic ANTHROPIC-COMPATIBLE backend override (DeepSeek, Mimo/Xiaomi, a
# self-hosted gateway, ...). That field still runs the SAME
# ``claude-agent-sdk``, just pointed at a different ``base_url`` via an API
# key instead of Anthropic OAuth (see ``_provider_registry.PROVIDERS`` /
# ``runtimes/_apptainer_provider.py``).
#
# :data:`AgentProvider` is a DIFFERENT axis entirely: ``spec.provider``
# (top-level, sibling of ``spec.runtime`` — see ``config._types.AgentConfig``)
# selects WHICH AGENT SDK / tool-calling framework backs the session at all —
# ``anthropic`` (``claude-agent-sdk``, talking to Anthropic OR an Anthropic-
# compatible gateway) vs ``openai`` (the ``openai-agents`` SDK, talking to
# OpenAI's Agents/Responses API). The two axes COMPOSE, they don't collide in
# practice — e.g. ``provider: anthropic`` (top-level, default) +
# ``claude.provider: mimo`` (nested) runs the Claude Agent SDK against
# Xiaomi's Anthropic-compatible endpoint. Do not conflate the two fields when
# reading/writing either one.
#
# Landed foundation-only: the schema + dataclass field exist and validate,
# but nothing branches on ``AgentConfig.provider`` yet — selecting
# ``"openai"`` today only affects `sac agents explain` / config introspection.
# openai-compat-2 lands the concrete ``openai`` runner (see
# ``_runners/_provider_session.py`` for the ``ProviderSession`` Protocol it
# will implement); openai-compat-3 wires the selection into the apptainer
# entrypoint + container image.
AgentProvider = Literal["anthropic", "openai"]

DEFAULT_AGENT_PROVIDER: AgentProvider = "anthropic"


[docs] @dataclass class ProviderSpec: """Vendor-agnostic backend override for the Claude SDK session. When set under ``spec.claude.provider``, the agent's SDK session talks to an Anthropic-SDK-compatible backend OTHER than Anthropic (DeepSeek, a self-hosted gateway, etc.) using a never-expiring, login-free API key instead of Anthropic OAuth. Lets bulk fleet work run on a cheap backend without burning Max-plan quota. The runtime injects three env vars into the container at start (see ``runtimes/_apptainer_provider.py``): * ``ANTHROPIC_BASE_URL`` ← :attr:`base_url` * ``SAC_ANTHROPIC_API_KEY`` ← the host value of ``$<auth_token_env>`` (bridged to ``ANTHROPIC_API_KEY`` for the SDK by sac's existing auth handoff). Fails loud at start if the env var is unset. * ``CLAUDE_CONFIG_DIR`` ← a clean per-agent dir — the conflict-breaker so the OAuth ``.credentials.json`` bind cannot win (apptainer ``--env`` is last-wins). The model id is the provider's own (e.g. ``deepseek-chat``); the claude-* regex in :mod:`config._validation` is relaxed whenever a provider is declared. Mutually exclusive with :attr:`ClaudeSpec.account` (OAuth pin) — an API-key backend needs no OAuth, so declaring both is a config error the runtime rejects loudly. """ base_url: str = "" """Anthropic-compatible base URL, e.g. ``https://api.deepseek.com/anthropic``. Required when the provider block is present.""" auth_token_env: str = "" """NAME of the host env var holding the API key (e.g. ``DEEPSEEK_API_KEY``) — NOT the key value. The operator sources the secret file; sac reads the env var's value at start and never logs it. Required when the provider block is present.""" allowed_tools: list[str] = field(default_factory=list) """Whitelist of built-in Claude Code tools to REGISTER when this provider is active. Maps directly to ``ClaudeAgentOptions.tools`` (the SDK ``--tools <csv>`` knob, see ``claude_agent_sdk/_internal/transport/subprocess_cli.py:241-250``), which controls the REGISTERED tool set rather than the use-allow list (``--allowedTools`` / ``--disallowedTools`` only affect permission, not whether the tool's JSON-schema ships in the outbound API request body — confirmed empirically by clew bm172 cohort v7 bind-test 2026-06-06). Why this matters: an Anthropic-API shim (LiteLLM / vLLM / gateway) may not recognize newer Claude Code built-ins (``ExitPlanMode``, ``BashOutput``, ``KillShell`` were added after LiteLLM 1.52.16). When the shim sees such a tool in the outbound ``tools[]`` it falls through its pydantic Union to the last subclass (``AnthropicComputerTool``), which requires ``display_width_px`` that the unknown tool's payload doesn't have → 422 on every API call → capsule errors all 60 turns. Whitelisting only the shim-recognized tools suppresses the unrecognized builtins at REGISTRATION, so they never enter the outbound body. Empty list (the default) means "use the runner's old-stable default whitelist when the provider is active, otherwise leave ``tools`` unset and let the CLI's full default register" — the runner picks the safe shape automatically. The string-form provider (e.g. ``provider: mimo``) doesn't carry this field today; the per-registry ``allowed_tools`` plumbing is a follow-up. Operators on the string form fall back to the runner's old-stable default until the registry plumbing lands. """